For Waste to Energy (WtE) plants, the decisions that protect availability are rarely made when an alarm sounds. The more useful moments sometimes come much earlier, when equipment behaves differently under comparable conditions: a gradual change in vibration, a shift in the relationship between temperature and load or a reading drifting from its established pattern.

Yet none of that automatically indicates a fault. Operating conditions may change, measurements are imperfect, and some failures often give little warning. The opportunity sits in the operational data your plant already holds, which can reveal a meaningful change in behaviour before an alarm condition is reached. The value to a plant engineer is not earlier detection, but credible, contextualised evidence while there is still time to decide whether anything needs investigating or doing differently.

Alarms and analytics provide different forms of visibility

Plant alarm systems are fundamental to safe operation. They tell you when a defined condition has reached a configured threshold, and action may be required. Predictive analytics offers a complementary view, considering whether a measurement is moving away from its established pattern, whether the relationship between variables is changing or whether equipment is behaving differently from before under comparable conditions.

A vibration reading can sit below its alarm threshold while steadily moving away from the equipment’s historical behaviour. A temperature can stay within an acceptable range while its relationship with load or pressure begins to change. Neither proves a failure is developing, but each is evidence that may justify a closer look. This does not replace plant controls, alarm management or operator experience; it adds visibility into changes that could become significant later.

What counts as normal depends on how the plant is operating

Recognising a meaningful change depends on knowing what normal behaviour looks like, and in a WtE plant that is rarely a single number. Plants run across changing loads, fuel characteristics, ambient conditions and operating strategies, and start-ups, shutdowns and maintenance all add legitimate variation. A value expected in one operating state may be unusual in another, so comparison against a simple average or static baseline can be misleading.

The signal often sits in the relationship between several measurements, the persistence or rate of a change, or the difference from earlier periods with comparable conditions. Those comparisons are hardest when deterioration develops slowly, because operators adapt to gradual change and today’s accepted picture can drift from how the equipment performed when new. A baseline that ignores load, plant state, recent maintenance and fuel conditions produces noise rather than insight.

An indication is not a diagnosis

Identifying unusual behaviour and diagnosing a fault are different tasks. Analytics can show that something is behaving differently from expectation, establish when the change began and highlight which other measurements are moving with it. That is still not a diagnosis.

Understanding the cause requires engineering interpretation, because an apparent anomaly could reflect a developing mechanical or process condition, or equally an operating change, different fuel, an instrumentation issue or a poorly represented operating state. As such, engineering judgement is part of how an observation becomes operationally meaningful.

Lead time is not the same as usable decision time

Discussions about predictive monitoring often focus on lead time: how long before an event a change was spotted. Lead time only matters when the indication is specific, timely and useful enough to support a decision. Three weeks of warning offers little if the indication is vague, reaches the wrong person or cannot be tied to a meaningful investigation. A much shorter warning can be worth far more if it gives your team time to verify the condition, involve the right expertise and prepare a response. Usable decision time is the more honest measure, because the point is not that the system noticed something earlier, but that the earlier notice allowed a better decision to be made.

What this looks like in practice

At an EcoPowerSoft customer Waste to Energy plant, the first sign of a developing boiler tube leak did not come from a conventional alarm. Late one evening, the machine learning model flagged two changes together: the difference between feedwater and steam flow was widening, and the make-up water supplied to the boiler was rising. Neither reading had crossed a protection threshold.

The plant team was notified the following morning and began to investigate. What confirmed the interpretation was not one measurement but several moving together. Feedwater flow kept climbing while steam flow held steady, flue gas moisture readings trended upward, and the economiser differential pressure rose as it began to foul, which increased ID fan load and moved the models watching the flue gas path into alert. The change was now visible across several systems, turning an early indication into a developing condition that the team could act on with confidence.

That evidence arrived six days before the unit had to come offline. Rather than managing a forced outage, the team used the time to locate the leak, plan a controlled shutdown and bring other outstanding work into the same window. The value was not a predicted failure date. It was six days of usable decision time, which for a plant measured on availability, is the difference between reacting and choosing how to respond.

Turning detection into a better decision

Detection is only half of it. An indication earns its value when there is a clear route from observation to review, with someone owning what happens next: who acts, how it is monitored when action is not yet justified and when it escalates. Volume is a poor measure here. A few well-contextualised observations that each warrant a decision beat a constant stream of alerts that do not.

The strongest applications of plant analytics start with an operational problem, not an ambition to deploy AI. Repeated causes of lost availability are an obvious place to look, along with anything costly to recover from or found later than you would like. The useful question is whether earlier awareness would change the response: whether it would let you inspect, gather evidence, bring in expertise, secure parts or make a different call.

There is one practical test for any predictive analytics application: what would we decide differently if we knew this sooner? If there’s no clear answer, detecting the change earlier may add little. If there is, the rest follows. Can the change be observed reliably, interpreted in context and put in front of the right person while there is still time to do something about it? That is when plant data stops being interesting and becomes operationally useful.

EcoPowerSoft works with Waste to Energy operators to turn the data a plant already holds into evidence teams can act on, so a developing change enters the decision-making process sooner. Adam Hanley, EcoPowerSoft’s Biomass and Energy from Waste Director, takes this further at Operational Optimisation 2026 on Tuesday 29th September, in his session “Driving Smarter Waste-to-Energy Performance with Connected Data”. If you’re attending, bring the operational problem you would most like earlier warning on, and Adam will talk through how it might be approached.